Xception Model 转移学习输入图像 - Tensorflow

Question

I am trying to use the xception model for a transfer learning task. I understand that it needs the minimum input shape of image to be (71, 71, 3) when downloading it with option include_top=False .

The issues I am facing is that when I try to reshape my data from (48,48,3) to (71,71,3) I am going in to RAM issues and my system gets restarted. Instead of reshaping the data outside I think of reshaping it within the network architecture. When I try to do that I have the follow error..

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-44-a448922440e7> in <module>()
      1 model = Sequential()
      2 #model.add(Input(shape=(48,48,3)))
----> 3 model.add(tf.keras.layers.Reshape((71,71,3), input_shape=(48,48,3)))
      4 #model.add(ReformatImage(71,71))
      5 model.add(conv_base)

8 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/layers/core.py in _fix_unknown_dimension(self, input_shape, output_shape)
    534       output_shape[unknown] = original // known
    535     elif original != known:
--> 536       raise ValueError(msg)
    537     return output_shape
    538 

ValueError: total size of new array must be unchanged, input_shape = [48, 48, 3], output_shape = [71, 71, 3]

my code is as below

from tensorflow.keras.applications import Xception
conv_base = Xception(weights='imagenet',include_top=False,input_shape=(71,71,3))

conv_base.trainable = False

model = Sequential()
#model.add(Input(shape=(48,48,3)))
model.add(tf.keras.layers.Reshape((71,71,3), input_shape=(48,48,3)))
#model.add(ReformatImage(71,71))
model.add(conv_base)
model.add(Flatten())
model.add(Dense(1024, activation='relu'))
model.add(Dense(1024, activation='relu'))
model.add(BatchNormalization())
model.add(Dense(10, activation='softmax'))

model.summary()

can someone please point me to right direction so that I can resolve this issue?

Answer 1

Reshape 'arranges' an input data from its inut shape to an output shape, so it doesn't create new data. Here you ask to arrange a (48x48x3) tensor into a (71x71x3) tensor, so it is not possible.

To do it you have to preprocess your images with resize functions: cv2.resize , skimage.transform.resize , tf.keras.preprocessing.ImageDataGenerator ...

Answer 2

You can use tf.keras.layers.experimental.preprocessing.Resizing :

tf.keras.layers.experimental.preprocessing.Resizing(48, 48)